With the advent of the Copernicus program and its wealth of open data, the Earth Observation domain is increasingly adopting big data technologies. This adoption is first made possible by efficient data storage and processing infrastructures, but most importantly by the development of data analytic applications with machine learning techniques. In this context, Thales Alenia Space (TAS) proposes to develop an application of change detection on Earth Observation satellite images time series. Examining changes of a designated area over a period of time, enables applications in many sectors, such as security, emergency, maritime and land surveillance. An important challenge in multi-temporal change detection is fast access and storage of a big amount of data and the computationally-intensive processing which is necessary. In this context, the use of the EVOLVE testbed, which contains HPC features, will be extremely useful. The objective of this pilot is to detect the changes over the entire Europe during one year.
TAS change detection tool provides generic change detection maps for pairs of time-consecutive Sentinel-2 data products that represent exactly the same field of view.
To do this, the first step consists in downloading Sentinel-2 data products from one of the Copernicus access platforms. The second step consists in transforming and pre-processing each Sentinel-2 data product in a single Sentinel-2 image with the bands of interest. The next step consists in sorting the Sentinel-2 images by time-series in chronological order to know on which pairs of images the resulting change map must be calculated. And the last step consists in computing the change detection maps in a smart way, based on a neural network model which allows to be more robust to the lighting and atmospheric condition differences between the two data acquisitions that do not represent changes of interest. (see figure below)
These resulting change detection maps are geoTiff rasters with the same size and geo-reference as the images on which they have been calculated, and whose pixel values range between 0 and 1 in order to represent the probability that a change has occurred between the acquisition dates of the two images.
This tool is implemented on Evolve platform and accessible through a Zeppelin notebook, which allows end users to make a request and launch the pipeline.
For each module (in black), a Docker image has been created and pushed to Evolve Docker registry. The different modules communicate between each other thanks to Kafka topics (in green), what allow to have coarse-grain parallelism. Argo Events (in orange) has been used to create Kubernetes pods from Kafka messages and release the GPU resource used by ChangeDetection module once the calculation is completed. And Dask scheduler has been used to have parallel computing inside the ChangeDetection module. All these technologies allow to mitigrate Thales Alenia Space tool in a cloud-native environment and considerably decrease the processing time.
Please, have a look at our “Change Detection pipeline on Evolve platform” video if you want more information.